Forge 2.0: Turn your documents into cited answers.
Forge 2.0 pairs sub-100ms dense retrieval with our in-house Voxell Reranker. Ask questions of your own documents and get the answer with the source attached, or test it live right now across 940,932 public passages from SEC filings, USPTO patents, and NASA reports.
POST /v1/answers/{corpus}/query → a cited answer in one call · API reference →The retrieval under all of this ranks #1 on MTEB, the public benchmark for search quality. Independently verified, and anyone can check the leaderboard. That is the sentence to repeat when someone asks if the answers can be trusted.
Our retrieval pipeline, measured on 7,817 real documents and 980,885 passages: the first result answers the question 83% of the time, and one of the top three does 90% of the time. 800 model-written questions across SEC filings, arXiv technical papers, NASA technical reports and USPTO patents, an easier test than one written by people; a model judge decides what counts as an answer. Read the receipts and the method →
Forge 2.0 is live. Follow the release stream.
Versioned releases keep our engineering accountable and give you measured capabilities to test every week. Start with Forge 2.0 today and see what ships next.
Two-stage retrieval with the Voxell Reranker and public 200-question receipts.
Single-stage cosine search misses critical clauses when dozens of passages share vocabulary. Forge 2.0 sweeps top-50 dense candidates at 87 ms P50 embedding latency with document-identity routing, then rereads all 50 candidates against your full question with our in-house Voxell Reranker and collapses duplicate passages.
- Live public workbench: Query 940,932 full-text passages across SEC filings, granted USPTO patents, and NASA technical reports with zero login.
- Signed 200-question receipts: 0.841 nDCG@10 on SEC filings (91% Answer@1), 0.787 on USPTO patents (86.5% Answer@1), 0.730 on NASA reports, and 0.900 on arXiv papers.
- Direct source verification: Every retrieved passage links straight to its official SEC EDGAR filing, Google Patents grant, or NASA NTRS record.
Start from where you actually are.
The same retrieval engine, three different front doors. Pick the one that sounds like your week.
One call. A cited answer comes back.
No vector database to provision, no chunking pipeline to babysit. Keep your stack; swap one URL.
- First real call in under five minutes, key issued at signup.
- OpenAI-compatible embeddings if you only want the model.
- Real prices on one page. No sales call anywhere.
Upload the docs. Share one link.
A private answer page for your team, set up by you, alone, today. No IT ticket, no code, nothing to install.
- Every answer cites its source, so you can defend it without rereading the originals.
- A corpus health score you can show a skeptic.
- Your team's name on it, not ours.
One vendor both halves will use.
Your engineers get the API. Everyone else gets no-code Spaces. One bill, one admin view, one security review.
- SSO and central administration on team plans; SAML, SCIM, and audit logs on Enterprise.
- A security page that states what we have and what we don't, in writing.
- Documents that can't leave your control? Forge Local runs the same engine inside your own AWS account.
- A corpus quality report you can put in front of a VP at day 90.
The part your security review will ask about.
Not stored, not trained on.
Embedding request text is not retained after the response, on every plan. Your content is not used to train models. Stated plainly in the security page and terms.
What we don't have, in writing.
No SOC 2 yet, and we say so before you ask. The Security and Trust page lists encryption, sub-processors, retention, and the gaps, so your review starts with answers instead of emails.
Or keep it in your account entirely.
Forge Local runs the same engine on your own AWS hardware. Nothing crosses our infrastructure, which is usually the shortest path through a review.
Already built RAG, and the answers are still wrong?
Send 10 to 50 failing queries and a sanitized slice of your corpus. We return your Recall@10, your nDCG@10, and the answers you are missing, measured, not guessed. Then you decide.
Free to start. Both doors open.
Try it on one real corpus before anyone asks for a card.